A Textile Standardization Large Language Model Integrating LoRA Fine-Tuning and Retrieval-Augmented Generation

Intelligent question answering for textile standards requires accurate domain understanding, clause retrieval, and evidence-grounded generation, yet general-purpose large language models often lack sufficient domain adaptation and factual support. This study proposes a textile-standard question-answering framework that integrates Low-Rank Adaptation (LoRA) fine-tuning with multidimensional retrieval-augmented generation (MDRAG). Based on 7017 domestic and international textile standards, we constructed a structured knowledge resource containing 121,166 annotated domain elements and a task-specific dataset of 2000 question–answer pairs. Qwen3.5-9B-Base, Llama 3.1-8B-Instruct, and GLM-4-9B were evaluated under three primary settings: Baseline, Baseline+LoRA, and Baseline+LoRA+MDRAG. A separate comparison with conventional RAG was conducted to assess the additional benefit of multidimensional retrieval over single-path retrieval. Among the primary settings, Qwen3.5-9B-Base with LoRA+MDRAG achieved the highest overall performance, with a cosine similarity of 0.624, a BLEU score of 0.283, and an expert-based overall score of 24.4. These findings support the use of lightweight domain adaptation together with multidimensional structured knowledge retrieval for textile-standard question answering.